Sensor hiperespectral para a predição de nitrogênio foliar, pigmentos e fotossíntese líquida na cultura do milho​

Abstract

The prediction of leaf nitrogen (NF) content, pigments and net photosynthesis (A) in maize leaves using hyperspectral data and machine learning (ML) models could become an essential tool in the rapid diagnosis of the behavior of these pigments in the face of nitrogen fertilization. The aim was to evaluate the accuracy of ML algorithms in predicting N content, pigments and photosynthetic rate in maize plants, using hyperspectral data as a basis. The experiment was carried out at the Federal University of Mato Grosso do Sul, Câmpus Chapadão do Sul in the second harvest of 2023. The experimental design was randomized blocks, in which the treatments consisted of four doses of N: N1 (0%), N2 (30% - 54 kg ha-1 de N), N3 (60% - 108 kg ha-1 de N) and N4 (120% - 216 kg ha-1 de N). At the V6 stage, spectral analysis was carried out on six leaf samples from each plot using a spectroradiometer= providing the 350 to 2500 nm bands. Once the wavelengths had been obtained, they were grouped into averages of representative band intervals. In addition to determining N and pigment content, photosynthesis analysis was carried out. The data was submitted to ML models: REPTree (DT), M5P Decision Tree (M5P), Random Forest (RF), Artificial Neural Networks (ANN), Support Vector Machine (SVM) and ZeroR (ZR) (control). Two input configurations were tested: using only the wavelengths to predict the variables (ALL) and using the spectral bands (SB). The following were used as output variables: NF content, pigments (chlorophyll and carotenoids) and the physiological variable A. Pearson's correlation coefficient (r) and mean absolute error (MAE) were used to assess the accuracy of the algorithms. The SVM algorithm obtained the best performance with the highest correlation (r) and lowest mean absolute error (MAE) between predicted and observed values for all variables except carotenoids, with the value of the correlation coefficient (r) above 0.6, and the error below 3.0. The M5P decision tree efficiently predicts all variables except chlorophyll a, achieving an r value of around 0.6 and an error below 0.3. The RF algorithm was able to predict the content of N, carotenoids and A, with an r-value of around 0.6. These three models performed best in predicting the variables using hyperspectral band data. Thus, the SVM, M5P decision tree and RF algorithms were effective in predicting NF, pigment and A content in maize, especially when SB was used as input data.

Keywords: Machine Learning. Gas Exchange. Support Vector Machine. Remote Sensing. Zea mays L.